Deep Noir framework autonomously steers LLMs, revealing new vulnerabilities

Researchers have developed a new framework called Deep Noir that automates the process of modifying Large Language Model (LLM) behavior at inference time. This method uses Logit Lens convergence and causal head-level attribution to autonomously discover optimal steering parameters. Across various model scales and architectures, Deep Noir has demonstrated significant improvements on tasks such as spam detection and sentiment analysis, achieving gains of up to 42 percentage points. The framework also revealed that steering interventions create a predictable prompt-injection attack surface, with vulnerability increasing alongside steering magnitude. AI

影响 Automates LLM behavior modification and reveals new security vulnerabilities for agent systems.

排序理由 Academic paper detailing a new method for LLM steering. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

• arXiv
• Deep Noir
• Hugging Face
• Logit Lens
• SST-2 Benchmark
• Transformer Models

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →